How Agents Evolve Without Degrading: From Risk Control to Semantic Engineering

01 – From "Can It Be Used" to "Can It Be Used Affordably, Stably, Long‑Term"

The hosts framed the conversation around three practical enterprise questions: collaboration, evolution, and cost control. Rather than treating agents as chat interfaces, they positioned them as tools that must integrate into real workflows, emphasizing stability, explainability, and traceability.

02 – First‑Principles of Architecture: Do Not Push Large Models Directly to the Decision Frontier

Li Qin argued that in banking, strict auditability demands that large language models (LLMs) not serve as the final decision maker. Instead, rule engines, fine‑tuned small models, and other deterministic modules guard the bottom line, while LLMs act as complex reasoners and evidence organizers – a "small‑model‑guards‑bottom‑line, large‑model‑extends‑capability" approach.

Item Qiao‑rui added a platform perspective: high‑certainty scenarios use workflows and rule chains combined with small models; exploratory tasks (dynamic analysis, attribution, news understanding) delegate to ReAct‑style agents that call tools autonomously. The underlying data, permission, and semantic layers are unified, keeping the agent lightweight while exposing atomic tools.

Zhao Heng emphasized designing verifiable sub‑tasks and a native function‑calling architecture that reduces reliance on heavyweight orchestration, allowing the system to benefit from model improvements without excessive complexity.

03 – When and Why Teams Choose to Abandon a Technical Route

The panel discussed abandoning direct LLM involvement in financial approval because of latency, hallucinations, and lack of explainability. They shifted the LLM role from "final decision maker" to "assistant for reasoning".

They also rejected the notion of a single all‑covering agent; instead, multiple specialized agents preserve clear business boundaries and simplify conflict resolution.

Further, they removed heavy third‑party frameworks, favoring modular, lightweight components that can be recombined, thereby reducing technical debt and improving flexibility.

04 – Measuring Agent Quality Beyond Correct Answers

Evaluation should focus on business‑level metrics (approval pass rate, bad‑debt reduction, complaint decline) rather than isolated task success.

Item highlighted the importance of "trajectory length": a short, low‑token, low‑tool‑call path indicates scalability, whereas long, costly traces suggest inefficiency.

Zhao stressed the need for benchmark suites to pinpoint failures in model reasoning, context handling, tool usage, or validation loops, especially to catch subtle hallucinations that appear correct.

05 – Stability Over Functionality

Multi‑step tasks amplify errors exponentially; small per‑step success rates can still collapse when chained.

Prompt engineering is described as "quasi‑mystical" because minor wording changes can cause large output variations, which is unacceptable in high‑risk finance.

Zhao added that user adoption hinges on clear onboarding; users must learn which decisions to trust the AI with and which to retain human control.

06 – Responsibility Allocation, Not Intelligence Comparison

The consensus is that responsibility, not raw capability, defines the human‑agent boundary. Humans set goals, calibrate processes, and accept final accountability, while agents handle high‑frequency execution and data organization.

Human‑in‑the‑loop remains essential because many scenarios carry legal and ethical consequences that cannot be fully automated.

07 – Budget Constraints: Save on Waste, Not on Model Size

Cost‑effective design allocates small models or rule systems to high‑frequency, deterministic steps, reserving large models for complex reasoning and cross‑modal analysis.

Zhao noted that overall cost overruns stem from overly long execution traces, poor context management, and unnecessary tool calls, not merely expensive model invocations.

Item emphasized atomizing capabilities into fine‑grained tools, which reduces token usage and execution time.

08 – From Pilot to Scale: New Problems Emerge

Scaling introduces challenges in maintaining performance across diverse organizations, data conditions, and user groups.

Zhao observed that increasing agent count raises coordination complexity, blurs boundaries, and amplifies responsibility conflicts, while also creating organizational resistance.

Li highlighted additional data silos, compliance constraints, and cross‑institution collaboration hurdles that appear only at scale.

Item stressed the need for user education so that users gradually learn when to delegate to AI and when to intervene.

09 – Future Opportunities and Uncertainties (1‑2 Years)

Experts anticipate deeper integration of reinforcement learning loops, private‑model deployment, and multimodal capabilities, especially in data‑engineered, risk‑controlled domains.

However, regulatory definitions of AI decision‑making, consumer appeal processes, and internal trust in AI versus job displacement remain the biggest unknowns.

10 – Conclusion: Agents as Engineering Tools for Decision Quality

The final takeaway is that agents are not meant to replace humans but to redistribute judgment, execution, and responsibility, enabling organizations to stay stable, clear, and evolvable amid growing complexity.